The world’s fastest Intel Alchemist GPU runs at 3586 MHz with the help of liquid nitrogen and the ElmorLabs EVC2.

https://www.youtube.com/watch?v=uH3f7iiKXqc

I also take P1 on all the available 3DMarks and will share all the tricks you need to know for a successful Intel Arc LN2 session.  

Let’s jump in!

Alchemist XOC Preparations

In preparation for this article, I finished a SkatterBencher guide demonstrating how to overclock this Intel Arc A770 GPU. At the end of the blog post, I said that we figured out workarounds for all the artificial limitations preventing us from maximizing the performance. Finding the workarounds was a requirement to even begin considering extreme overclocking.

I also want to give a quick shout-out to Reddit user Un8ounded. I thought I’d be the first to try liquid nitrogen on an Intel Arc card, but he did it first and shared the experience at length on Reddit about 2 months ago.

After finishing the SkatterBencher overclocking guide, I pinged Elmor from ElmorLabs to see if I could spend a day at his office for an LN2 overclocking session.

Why Overclock Alchemist With Liquid Nitrogen?

I’m sure you’re thinking: why would anyone want to overclock Intel Alchemist?

Aside from the fact that I’m a SkatterBencher and wish to uncover every chip’s performance potential, Arc is Intel’s first foray into discrete graphics in a long time. So, it’s interesting to see how their design scales unconstrained by temperature and voltage.

In addition, it’s also interesting to see how the TSMC N6 process scales with elevated voltages and subzero temperatures. Apart from Intel Arc, the only discrete graphics cards using TSMC N6 are the Radeon RX 6500 XT and RX 6400. I did overclock the former in SkatterBencher #41, but due to AMD’s poor support for GPU overclocking, it’s impossible to see how fast they can go.

What Are My Objectives?

My main objective was to have a legitimate claim to the World’s Fastest Intel Alchemist GPU title. I figured I needed two data points for that: one, obviously, to achieve the highest frequency ever on Alchemist, and two, to achieve a competitive benchmark P1.

There isn’t a great resource to reference for the highest GPU frequency. So, I picked two arbitrary targets. The first target is to get the highest frequency in GPU-Z regardless of the load. The second target is achieving the highest frequency in a light workload. I picked Furmark GPU Stress Test with 50×50 pixels as the light workload.

For the benchmark objective, I downloaded all ten available 3DMark workloads and sorted them by number of submissions with the Arc A770 graphics card. My main objective was to achieve P1 in the top-3 most used benchmarks, but if there’s time and nitrogen left, I’d also try the others.

The most used benchmark with the Arc A770 is Time Spy, and the target score to beat was 16,415 points by Brutuscat2, who had the card run an average frequency of 2800 MHz.

Hardware Setup

Next is preparing the setup. This immediately got much more complicated than I wanted.

The usual process for GPU overclocking involves picking up a GPU LN2 container like the Kingpincooling TEK-9. Unfortunately, the Arc A770 has quite odd mounting holes, and there’s no mounting bracket available for any commonly available GPU LN2 container.

The next best thing for extreme GPU overclocking is using a PCIe riser to place the GPU horizontally. Fortunately, Elmor had an interesting 90-degree PCIe riser at his office, making it extremely easy to lay the GPU flat and use the ElmorLabs Volcano LN2 container.

Using a CPU container does come with additional risk because it’s so heavy and could crack the GPU die if placed poorly. To protect the GPU die, I added two layers of protective tape around the die.

Additionally, using a CPU cooler isn’t ideal because it has no GPU mounting bracket. So, we must rely on mother nature’s mounting force: gravity.

The last thing to sort out is the insulation around the PCIe area. Basic insulation sufficed because I wasn’t planning on running the card for too long and cold.

In addition to the Volcanoo LN2 container, I also used other ElmorLabs gear. I use the EVC2SE to program the voltage controller over the I2C interface. I also use the ElmorLabs KTH-USB to read out the LN2 container temperature and monitor it using the software.

Working Around The Limitations

The point of this video is not just to show my results but also to share solutions to the artificial overclocking limitations of the A770. There are five noteworthy limitations:

1. Fixed OC Frequency
2. Power Limit
3. Voltage Limit
4. Voltage Throttle
5. VRM Temperature Throttle

I cover these limitations in my SkatterBencher overclocking guide at length, so check out that if you want to learn more. I’ll try to keep the explanations as brief and concise as possible.

Fixed OC Frequency

By default, the Arc GPUs rely on a factory-fused voltage-frequency curve to adjust the boost frequency dynamically. However, we want to precisely control the effective GPU frequency during extreme overclocking sessions. Unfortunately, those controls are not available in the Arc Control software.

In the summer of 2022, when the Arc A-Series GPUs had just become available, there weren’t many third-party overclocking tools. With the help of Shamino, we built the Arc OC Tool, which offers a simple interface to access GPU overclocking toolkit available via the publicly available Intel Graphics Control Library (IGCL).

A vital benefit of the Arc OC tool is exposing the OC Lock feature, which is unavailable in the Arc Control software. With OC Lock, you can set a specific operating voltage and frequency. In addition, the OC Lock function isn’t limited to a maximum 300 MHz frequency offset offered in Arc Control. So, we can set the frequency as high as we need to.

Power Limit

The 228W total board power limit is the first bottleneck when overclocking the Intel Arc A770, even on ambient cooling. The power limit on the A770 is the total board power – meaning all the power of all components on the graphics card.

The GPU manages the power limit in two steps. First, the voltage controller provides information about the output power to the GPU, then the GPU enforces the limit through software. You can override the power limit in two ways:

1. Use the Predator Bifrost software application and trick the profiles into applying higher power limits. This is a neat trick shared by ElmorLabs discord user Dyno.
2. Program the voltage controller to report lower than actual power consumption to the GPU

Predator BiFrost Power Limit Trick

Here’s the process to increase the power limit using the Predator BiFrost software. If you have trouble finding the software, I uploaded a mirror here: Application_Acer_1.0.104.0_W11x64_A.zip.

1. Open the Predator Bifrost software, create a new profile, then close the software.
2. Now browse to AppData\Roaming\PredatorBifrost\Presets and open the settings file.
3. Now look for your custom profile and change the PowerLimit value to anything you want. Make sure to save the file.
4. Open the Acer Predator Bifrost software again and activate your custom profile.
5. You can verify with HWiNFO that the power limits were set correctly.

MP2979 Power Limit Trick

To determine the method used on the Arc A770, we must look closer at the voltage regulator design.

The GPU is powered by a 6-phase design managed by Monolithic Power Systems (MPS) MP2979 digital multi-phase voltage controller. This controller drives six (6) Monolithic MP86956 70A Intelli-Phase DrMOS, one for each phase.

The hardware modification adds an I2C pin header on the graphics card PCB, allowing us to communicate directly with the onboard digital voltage controllers. We can then connect the EVC2SE device to the I2C pins to control the voltage regulator.

The relevant function to adjust the power limit is called Iout Gain. By lowering the programmed value, this function allows us to skew the reported output current by a specific factor. So, we can use it to have the voltage controller report to the GPU that it’s using much lower power than it actually uses.

Voltage Limit & Throttle

The voltage limit and throttle are separate issues but can be resolved with one solution.

When we set the voltage using OC Lock, the set voltage is offset by about +215mV. So, if you set 1V in the software, the actual voltage would be around 1.215V. In addition, there’s a limit to how high you can program the voltage. Any OC Lock value over 1.1V (or about 1.295V actual voltage) doesn’t work. So, we’re limited to a maximum voltage of 1.295V.

Next, the GPU also imposes an artificial voltage performance throttler as the GPU automatically reduces the operating frequency when the voltage exceeds a certain threshold. On the A770, this GPU frequency throttling mechanism kicks in when the set voltage is over 1.2V.

We can solve both issues by switching the voltage regulator control mode from SVID to PMBus.

• In SVID mode, the voltage controller adjusts the output voltage based on the GPU request. The GPU relies on its factory-fused voltage-frequency curve to determine the appropriate voltage at any moment.
• In PMBus mode, we program the voltage controller output voltage directly without interference from the GPU. Thus, we can set any voltage over 1.2V.

VRM Temperature Throttle

The last performance limiter on this card is the VRM temperature. As I highlight in my Arc A770 overclocking guide, the VRM design with 6 power stages gets in real trouble when the power consumption exceeds 330W. When the VRM temperature hits 110 degrees Celsius, it reduces the GPU clock to its base frequency of 2100 MHz. This won’t be great when we use higher voltage with liquid nitrogen.

The Junction Temperature of the DrMOS power stages is 150 degrees Celsius, and the voltage controller’s over-temperature protection (OTP, F2h) is set accordingly at 150 degrees Celsius. However, the voltage controller’s maximum temperature (F3h) is set to 115 degrees Celsius. When this temperature is reached, the VRHOT# asserts.

VR_HOT# is one of many protections safeguarding your graphics card. It is a required signal from the voltage regulator to the chip, informing the chip that the VRM temperature is too high. In that case, the chip will throttle its frequency to ensure protection. So, we can provide additional thermal headroom by adjusting the voltage controller’s VR_HOT threshold.

Alternatively, you can design a better thermal solution for your card’s VRM. Maybe you can even use a liquid nitrogen container, just like South African overclocker Vivi did to qualify for the MSI MOA 2013 Finals.

Challenges and Expectations

Before we get to the results, let me quickly go over some of the challenges and expectations for the system.

The first and most obvious challenge will be the VRM capability. As I highlighted in my Arc A770 SkatterBencher guide: the 6 DrMOS power stages get in trouble when the card’s using over 330W. On water cooling, that happens when the voltage is around 1V. Lower temperatures also lower the power consumption at the same voltage due to reduced leakage current, but we know that the VRM will become a limitation at some point.

When it comes to cold bug – that’s the temperature point the card stops operating – we are in uncharted territory. So, we have no idea what the GPU, or the GDDR6 memory, can handle. For what it’s worth, many extreme overclockers nowadays use GPU heaters to keep the memory at positive temperatures, as they’re incapable of running at low temperatures.

Alchemist XOC Results

The overclocking session started well as we were running 3DMark Time Spy within an hour at 3 GHz. However, the 3DMark Time Spy benchmark performance was poor, scoring less than 16,000 points. That’s miles away from our target. The performance of 3DMark Port Royal, a ray-tracing benchmark, was even worse: only 7500 points, far from the 8200 points we needed for P1.

What was going on?

In hindsight, the poor performance result had an obvious root cause. Maybe you can spot it? But as is typical for an extreme overclocking session, it took hours to figure out what happened. So, what was the root cause? The GPU was running at PCIe Gen 4 x1. We could’ve spotted that in GPU-Z and eventually confirmed it by checking the BIOS. Turns out that this particular CPU had some PCIe problems. After swapping in another Raptor Lake CPU, we could finally run the GPU at Gen 4 x16. After that, the good results came in fast.

So, let’s get on to the final results. I have many results to show, so I’ll go in order of GPU frequency.

The highest frequency I achieved with Alchemist was 3586 MHz with 1.305V. The LN2 container temperature was -80 degrees Celsius, and the GPU temperature was about -50 degrees Celsius. The card was very unstable at this point. While I could idle a couple of seconds at this frequency, eventually, the screen would lock up, and a reboot was required.

The highest frequency during a light workload was 3400 MHz with 1.285V. The LN2 container temperature was -70 degrees Celsius, and the GPU temperature was about -30 degrees Celsius. Despite running Furmark GPU Stress Test with only 50×50 pixels, the power consumption was more than 170W.

The highest frequency during a 3DMark benchmark was 3100 MHz, though unfortunately, I could never complete more than 1 Time Spy Game Test. Well, that’s not entirely true. I was able to complete a Time Spy benchmark once, but I had forgotten to extend the VRM temperature limit. So, during the benchmark, the card sometimes throttled to 2100 MHz when the VRM hit 110 degrees Celsius resulting in a poor benchmark result.

The highest frequency to complete a 3DMark benchmark was 3012 MHz with 1.12V. The container temperature was about -50 degrees Celsius, and the GPU temperature was -30 degrees Celsius. I needed this frequency to get P1 in the 3DMark Night Raid benchmark.

Speaking of the 3DMark benchmarks, I managed to sweep P1 in all but one of the ten 3DMark benchmarks. I also achieved this with a 3 GHz GPU clock, making it a nice, neat milestone.

Here are the links to my 3DMark results:

• Fire Strike https://www.3dmark.com/fs/30414023
• Fire Strike Extreme https://www.3dmark.com/fs/30413944
• Fire Strike Ultra https://www.3dmark.com/fs/30413973
• Night Raid https://www.3dmark.com/nr/935001
• Port Royal https://www.3dmark.com/pr/2458906
• Speed Way https://www.3dmark.com/sw/718383
• Speed Way Stress Test https://www.3dmark.com/swst/364653
• Time Spy https://www.3dmark.com/spy/40252999
• Time Spy Extreme https://www.3dmark.com/spy/40253840
• Wild Life https://www.3dmark.com/wl/312567
• Wild Life Extreme https://www.3dmark.com/wl/312566

It may surprise you, but I also ran (and passed!) the Speedway Stress Test at 3GHz to learn the limits of this GPU. The GPU voltage was 1.114V. The GPU temperature was -20 degrees Celsius with an LN2 container base temperature of -44 degrees Celsius. The VRM temperature was 85 degrees Celsius, and, surprisingly, the memory temperature was 71 degrees Celsius. The total GPU power was 258.404 watts.

Lastly, I tried running the GPU at -90 degrees Celsius. Aside from the GDDR6 temperature-related instability preventing me from running any benchmarks, it seemed the card had no qualms about this low temperature.

Discussion & What’s Next

So, I took P1 in all but one benchmark – is there anything left in the tank? Yes, I think so.

As I mentioned several times, the VRM is the main bottleneck for this card. Other Arc A770 cards on the market with an 8-phase VRM design may be better suited for chasing performance records. Additionally, there’s always the option to zombify the graphics card by soldering an external VRM to the card.

The GPU would also benefit from better mounting, and it shouldn’t be challenging to do better than this gravity-reliant mechanism. You can control the GPU temperature better and maybe finetune for the optimal voltage with a better mounting.

Lastly, I’m pretty mediocre at GPU extreme overclocking … I’m sure plenty of people out there can push the card to its limits much better than I can. I hope by sharing all the information in this video, those with more talent now have the tools to push Intel Alchemist to its limit.

Conclusive Thoughts

I think it’s already clear from the video, but I’m delighted with the outcome of this project.

The first significant milestone was gathering all the information and tools to work around the various artificial limitations. I did that in my SkatterBencher guide. Check!

The second important milestone was to find out what Intel Alchemist is capable of. Intel’s Alchemist hitting a maximum frequency of over 3.5 GHz is impressive for a first stab at modern GPUs. I remember when the industry raved about AMD’s 3.3 GHz with RDNA2 two years ago. In that light, we can’t be anything else than impressed by Alchemist’s capabilities.

Now that all the information and tools are out there, I can’t wait to see what experienced extreme overclockers can do with these GPUs.

To end the video, I want to say a big thanks to ElmorLabs for the hospitality. Of course, I also want to thank you for reading and the Patreon supporters for supporting my work. If you have any questions or comments, please drop them in the comment section below. 

Until the next time!